AI Engineer · LLM Agents · Voice AI · Testing & Evaluation
I build LLM agents and voice AI products, and I measure them before I trust them. Right now I'm building Katha, an iOS reader that gives every character in a book its own AI voice. Before that I launched indes.ai, a multi-LLM platform with agentic routing.
Katha: Multi-Voice AI Reader for iOS · in development
- Agentic voice casting: Claude profiles each character from the book and writes a voice prompt; Qwen3-TTS renders candidates until one passes a pitch check, then it becomes that character's voice
- Speaker attribution (NER, alias resolution, Claude as LLM fallback), measured on a 380-quote annotated benchmark: explicit-quote accuracy 0.68 → 0.82 at 95% precision
- 4 TTS engines routed by language: Kokoro on-device (MLX), self-hosted Qwen3-TTS and Indic Parler-TTS, Apple as fallback, with consent-gated voice cloning
- 1,500+ Swift tests and 150+ Pytest tests in CI · Swift, Python, Claude API, MLX
indes.ai: Multi-LLM Platform with Agentic Routing
- Streams GPT, Claude, Gemini and DeepSeek side by side and merges them into one answer (web, iOS/Android)
- Agentic routing: 11-type query classifier with LLM fallback, plus a 7-tool orchestrator that answers simple queries without an LLM call
- Fine-tuned a T5 synthesis model on Claude outputs, evaluated with BLEU/ROUGE
- 650+ Jest/Supertest tests gating every PR, Playwright E2E across Chromium, Firefox and WebKit · Node.js, React Native, Python
aria: LLM red-teaming agent
- 10 attack strategies, 77 variants
- Reflexion-based learning from failed attempts
reflexive-llm-agent: self-reflective tool-using agent
- ReAct + Reflexion loop with long-term ChromaDB memory
- FastAPI backend and Streamlit UI
- vllm-project/llm-compressor#2330: GSM8K evaluation script and AWQ + FP8 results (merged)
- instructlab/training#682: fix HYBRID_SHARD failure when world_size < available GPUs (merged)
- triton-kernels: FlashAttention, MatMul, Softmax and LayerNorm in Triton, autotuned for T4/V100/A100/H100
- ddp-lora-trainer: multi-GPU LoRA fine-tuning with PyTorch DDP and mixed precision
- llm-quantization · llm-sparsity: GPTQ, AWQ, SmoothQuant; magnitude, 2:4 and WANDA pruning
- vibeclean-ai: AST-based security analysis for AI-generated code
- cuda-mempool · lockfree-queue · arena-alloc: CUDA, Rust and C++ systems work
- Agents & LLMs: Claude, GPT, Gemini APIs · tool calling · agentic routing · RAG · ReAct · LangChain · MCP
- Voice AI: Qwen3-TTS · Indic Parler-TTS · Kokoro · Chatterbox · MLX
- Testing & Evaluation: Playwright · Jest · Pytest · Swift Testing · load testing · eval sets · BLEU/ROUGE
- Languages: Python · Swift · JavaScript/Node.js · SQL · C++ · Rust
- Infrastructure: Google Cloud · AWS · Docker · GitHub Actions · Redis · PostgreSQL · Kafka